By Devin Didericksen on January 7, 2015

   

A very good machine learning paper was published a few months ago where the authors trained a neural network to simulate the style of one picture on top of the content of another picture. Their results were pretty intriguing.
 

To illustrate the concept, if you trained a neural network to learn the painting style of Picasso, and you gave it a photo of Obama to “paint”, you’d get something back like this:
 


 

Likewise, here is the Eiffel Tower painted in the style of Van Gogh’s Starry Night:
 


 

This neural artwork capability was recently added to MXNet, which is a relatively new deep learning framework (from the same guys that made XGBoost). I’ve been learning it over the past month, and I have to say I love it so far. I decided to train my own neural networks to paint one of my photos in several different styles. The photo I used is this one I took last summer of the Colosseum:
 


 

The neural artwork results are below. My favorite is the first one, in the style of Edvard Munch’s “The Scream”. If you are interested in doing your own neural art I suggest following this tutorial.